MindCV is an open-source toolbox for computer vision research and development based on MindSpore. It collects a series of classic and SoTA vision models, such as ResNet and SwinTransformer, along with their pre-trained weights and training strategies. SoTA methods such as auto augmentation are also provided for performance improvement. With the decoupled module design, it is easy to apply or adapt MindCV to your own CV tasks.

>>> import mindcv
# create a dataset
>>> dataset = mindcv.create_dataset('cifar10', download=True)
# create a model
>>> network = mindcv.create_model('resnet50', pretrained=True)

# transfer learning in one command line
>>> !python train.py --model=swin_tiny --pretrained --opt=adamw --lr=0.001 --data_dir={data_dir}

The performance of the models trained with MindCV is summarized in benchmark_results.md, where the training recipes and weights are both available.

Model introduction and training details can be viewed in each subfolder under configs.

• mindspore >= 1.8.1
• numpy >= 1.17.0
• pyyaml >= 5.3
• tqdm
• openmpi 4.0.3 (for distributed mode)

To install the dependency, please run

pip install -r requirements.txt

MindSpore can be easily installed by following the official instructions where you can select your hardware platform for the best fit. To run in distributed mode, openmpi is required to install.

The following instructions assume the desired dependency is fulfilled.

The released version of MindCV can be installed via PyPI as follows:

pip install mindcv

The latest version of MindCV can be installed as follows:

pip install git+https://github.com/mindspore-lab/mindcv.git

Notes: MindCV can be installed on Linux and Mac but not on Windows currently.

To get started with MindCV, please see the transfer learning tutorial, which will give you a quick tour on each key component and the train/validate/predict pipelines.

Below are a few code snippets for your taste.

>>> import mindcv
# List and find a pretrained vision model
>>> mindcv.list_models("swin*", pretrained=True)
['swin_tiny']
# Create the model object
>>> network = mindcv.create_model('swin_tiny', pretrained=True)
# Validate its accuracy
>>> !python validate.py --model=swin_tiny --pretrained --dataset=imagenet --val_split=validation
{'Top_1_Accuracy': 0.808343989769821, 'Top_5_Accuracy': 0.9527253836317136, 'loss': 0.8474242982580839}

Image classification demo

Infer the input image with a pretrained SoTA model,

>>> !python infer.py --model=swin_tiny --image_path='./tutorials/data/test/dog/dog.jpg'
{'Labrador retriever': 0.5700152, 'golden retriever': 0.034551315, 'kelpie': 0.010108651, 'Chesapeake Bay retriever': 0.008229004, 'Walker hound, Walker foxhound': 0.007791956}

The top-1 prediction result is labrador retriever (拉布拉多犬), which is the breed of this cut dog.

It is easy to train your model on a standard or customized dataset using train.py, where the training strategy (e.g., augmentation, LR scheduling) can be configured with external arguments or a yaml config file.

• Standalone Training

# standalone training
python train.py --model=resnet50 --dataset=cifar10 --dataset_download

Above is an example for training ResNet50 on CIFAR10 dataset on a CPU/GPU/Ascend device

• Distributed Training

For large datasets like ImageNet, it is necessary to do training in distributed mode on multiple devices. This can be achieved with mpirun and parallel features supported by MindSpore.

# distributed training
# assume you have 4 GPUs/NPUs
mpirun -n 4 python train.py --distribute \
	--model=densenet121 --dataset=imagenet --data_dir=/path/to/imagenet

Notes: If the script is executed by the root user, the --allow-run-as-root parameter must be added to mpirun.

Detailed parameter definitions can be seen in config.py and checked by running `python train.py --help'.

To resume training, please set the --ckpt_path and --ckpt_save_dir arguments. The optimizer state including the learning rate of the last stopped epoch will also be recovered.

• Config and Training Strategy

You can configure your model and other components either by specifying external parameters or by writing a yaml config file. Here is an example of training using a preset yaml file.

mpirun --allow-run-as-root -n 4 python train.py -c configs/squeezenet/squeezenet_1.0_gpu.yaml

Pre-defined Training Strategies: We provide more than 20 training recipes that achieve SoTA results on ImageNet currently. Please look into the configs folder for details. Please feel free to adapt these training strategies to your own model for performance improvement， which can be easily done by modifying the yaml file.

• Train on ModelArts/OpenI Platform

To run training on the ModelArts or OpenI cloud platform:

1. Create a new training task on the cloud platform.
2. Add run parameter `config` and specify the path to the yaml config file on the website UI interface.
3. Add run parameter `enable_modelarts` and set True on the website UI interface.
4. Fill in other blanks on the website and launch the training task.

To evalute the model performance, please run validate.py

# validate a trained checkpoint
python validate.py --model=resnet50 --dataset=imagenet --data_dir=/path/to/data --ckpt_path=/path/to/model.ckpt

• Validation while Training

You can also track the validation accuracy during training by enabling the --val_while_train option.

python train.py --model=resnet50 --dataset=cifar10 \
		--val_while_train --val_split=test --val_interval=1

The training loss and validation accuracy for each epoch will be saved in {ckpt_save_dir}/results.log.

More examples about training and validation can be seen in examples/scripts.

• Graph Mode and Pynative Mode

By default, the training pipeline train.py is run in graph mode on MindSpore, which is optimized for efficiency and parallel computing with a compiled static graph. In contrast, pynative mode is optimized for flexibility and easy debugging. You may alter the parameter --mode to switch to pure pynative mode for debugging purpose.

Pynative mode with ms_function is a mixed mode for comprising flexibility and efficiency in MindSpore. To apply pynative mode with ms_function for training, please run train_with_func.py, e.g.,

python train_with_func.py --model=resnet50 --dataset=cifar10 --dataset_download  --epoch_size=10

Note: this is an experimental function under improvement. It is not stable on MindSpore 1.8.1 or earlier versions.

We provide the following jupyter notebook tutorials to help users learn to use MindCV.

• Learn about configs
• Inference with a pretrained model
• Finetune a pretrained model on custom datasets
• Customize your model //coming soon
• Optimizing performance for vision transformer //coming soon
• Deployment demo

Currently, MindCV supports the model families listed below. More models with pre-trained weights are under development and will be released soon.

• 2023/03/05

1. Add Lion (EvoLved Sign Momentum) optimizer from paper https://arxiv.org/abs/2302.06675
   • To replace adamw with lion, LR is usually 3-10x smaller, and weight decay is usually 3-10x larger than adamw.
2. Add 6 new models with training recipes and pretrained weights for
   • HRNet
   • SENet
   • GoogLeNet
   • Inception V3
   • Inception V4
   • Xception
3. Support gradient clip

• 2023/01/10

1. MindCV v0.1 released! It can be installed via PyPI pip install mindcv now.
2. Add training recipe and trained weights of googlenet, inception_v3, inception_v4, xception

• 2022/12/09

1. Support lr warmup for all lr scheduling algorithms besides cosine decay.
2. Add repeated augmentation, which can be enabled by setting --aug_repeats to be a value larger than 1 (typically, 3 or 4 is a common choice).
3. Add EMA.
4. Improve BCE loss to support mixup/cutmix.

• 2022/11/21

1. Add visualization for loss and acc curves
2. Support epochwise lr warmup cosine decay (previous is stepwise)

• 2022/11/09

1. Add 7 pretrained ViT models.
2. Add RandAugment augmentation.
3. Fix CutMix efficiency issue and CutMix and Mixup can be used together.
4. Fix lr plot and scheduling bug.

• 2022/10/12

1. Both BCE and CE loss now support class-weight config, label smoothing, and auxiliary logit input (for networks like inception).

• 2022/09/13

1. Add Adan optimizer (experimental)

We appreciate all kind of contributions including issues and PRs to make MindCV better.

Please refer to CONTRIBUTING.md for the contributing guideline. Please follow the Model Template and Guideline for contributing a model that fits the overall interface :)

This project follows the Apache License 2.0 open-source license.

MindCV is an open-source project jointly developed by the MindSpore team, Xidian University, and Xi'an Jiaotong University. Sincere thanks to all participating researchers and developers for their hard work on this project. We also acknowledge the computing resources provided by OpenI.

If you find this project useful in your research, please consider citing:

@misc{MindSpore Computer Vision 2022,
    title={{MindSpore Computer  Vision}:MindSpore Computer Vision Toolbox and Benchmark},
    author={MindSpore Vision Contributors},
    howpublished = {\url{https://github.com/mindspore-lab/mindcv/}},
    year={2022}
}